Orelon logoOrelon
料金

Cinematic AI Video for Regional Agencies: A Working Guide

2026年11月17日 · Orelon Team 著

AI動画テンプレートを見る

着想のためにコミュニティ作品をいくつか閲覧し、任意のテンプレートを開いて Orelon で作成を続けましょう。

A shot-by-shot workflow for regional studios that want cinematic AI video: briefing, prompting, consistency, sound, QA, delivery and pricing.

Every regional studio has had the same conversation. The client wants a film that looks like it cost six figures, the budget is a fraction of that, and the delivery date is three weeks out. AI video generation has made that conversation winnable, but only for teams that treat it as a production discipline rather than a slot machine.

This guide is for agencies, in-house marketers and freelancers working outside the big production centres, in towns where the local economy runs on independent businesses with real stories. What follows is a workflow: how to brief, how to prompt, how to keep a sequence coherent, how to finish, and how to price the work so it stays profitable. The model you choose matters far less than what you do before and after generation. That is where cinematic quality is actually decided.

What cinematic means when the camera is a prompt

Cinematic is a set of decisions, not a filter. Motivated light. A camera that moves for a reason. A world that stays consistent from shot to shot. Sound that carries the emotion the image cannot. And restraint, which is the confidence to hold a shot instead of cutting every second and a half.

A generative model hands you texture: grain, bokeh, weather, skin, atmosphere. It does not hand you intent. A bakery spot built from the prompt cozy bakery, warm light, cinematic will look like stock footage with extra haze. The same bakery, briefed as a flour-dusted hand pushes a tray into an oven at dawn, steam catching a single window light, camera locked off at knee height, looks like a film, because someone decided what the shot was for.

That decision should be written down before you open a generator. It is the cheapest step in the process and the one most often skipped. A one-page shot rationale, with a sentence per shot explaining why it exists, will save you more time than any prompt library you can collect.

Texture is free, intent is not

The cost of a mediocre frame has collapsed to almost nothing, which means mediocre frames are now everywhere. The scarce resource is judgement. When a client says the last agency's video felt cheap, they are rarely reacting to resolution or frame rate. They are reacting to footage that had no reason to be there: six shots of the same thing, a camera that drifts because drifting looks premium, music that never breathes.

Train yourself to ask one question of every shot: if I removed this, what would the viewer fail to understand? If the answer is nothing, cut it. That single habit will lift a regional agency's output above most of what its clients see online.

The division of labour: what AI should and should not touch

AI video is strongest on establishing shots and plates, texture inserts such as pouring, steam, fabric, sparks and water, stylised montages, animatics for pitching, and pickup shots you cannot justify returning to a location for. It is weakest on dialogue performance, complex hand interaction with objects, brand-accurate packaging, and anything where a real face must be recognisable and consistent across twenty cuts.

Split the shot list along that line. Generate what the tool is good at, shoot or license the rest. Clients do not care about the mix; they care that the finished piece holds together as one thing.

Shot type AI suitability Practical approach
Establishing plate, landscape, street High Generate three variations, grade to match
Texture insert, product detail High Macro prompt, hard top light, short duration
Presenter to camera Low Shoot on a phone with a lav mic if needed
Hands assembling a product Medium Generate wide and cut around the contact point
Logo end card None Build in the editor, never generate type

That last row matters. Generated text is a tell. Every logo, legal line and URL should be added in post from a clean asset, not coaxed out of a model.

Pre-production that decides the outcome

You still need a one-line brief, a reference board of eight to twelve stills, a locked aspect ratio list, a location description, wardrobe and prop notes, and a shot count that matches the runtime. Ten seconds of planning per finished second of film is a reasonable ratio: a 45-second piece usually needs five to seven shots plus inserts.

Reference frames are the biggest quality lever

Reference stills do double duty. They align the client, and they become image-to-video starting frames, which is the single most reliable way to hold composition and character today. Pull them from your own photography, from previous projects, or generate them first in an image tool. The AI image generator route is often faster than hunting for licensed stills, and it keeps the look coherent across a set.

Agree on the stills before you generate a single second of motion. A client who signs off on a frame rarely rejects the shot that grows from it. A client who has only read a description will reject almost anything.

The six fields of a shot row

Each row of the shot list should carry a shot number, a beat, a description, a camera note (lens, height, movement), a lighting note (source, direction, quality) and an audio intent. That last field is the one people forget. If you do not know whether a shot should carry room tone, a tool strike or a music swell, you do not yet know why the shot is in the film.

Prompting with photographic language

Vague adjectives produce vague footage. Concrete photographic terms produce repeatable results. Treat them as vocabulary, not magic words.

Intent Write this Avoid this
Intimacy 50mm, chest height, shallow depth of field, static emotional, close
Scale 24mm, low angle, slow dolly forward epic wide
Detail 100mm macro, top light, hard shadow close up detail
Movement handheld micro-drift, slow pan left dynamic
Mood single practical lamp, deep shadow falloff moody, dark

Name the source of light

Single window, camera left, soft, falling off into black is actionable. Golden hour is a cliché that produces the same stock sunset every time. For interiors, specify practicals: a desk lamp, an oven light, a fridge door, a phone screen. Practical sources give the model something to anchor reflections and shadows to, and reflections are what make footage feel photographed rather than rendered.

Motion deserves the same discipline. Slow moves read as expensive. Fast moves read as chaotic unless the shot is built for them. If you cannot describe the move in five words, the shot is probably two shots.

Build modular prompt blocks

Keep a reusable structure: subject and action, environment, camera, light, texture, exclusions. Paste the same camera and texture blocks across every shot in a scene so the look carries, and change only the subject and action lines. Negative guidance belongs in the same slot every time: if you do not want lens flares, watermarks or over-smoothed skin, say so consistently. A shared prompt library keeps a small team writing in the same voice instead of five different ones.

Consistency across a sequence

A 60-second commercial collapses the moment the protagonist's jacket changes colour between shots. Consistency is a bookkeeping problem before it is a technical problem.

Build a character sheet: front, three-quarter, profile, plus wardrobe and hair notes in text. Keep product references as clean cut-outs on neutral backgrounds. Name every asset predictably, for example filmname_sh03_hero_front_v2. When you have forty files, naming is the difference between a smooth edit and a lost afternoon.

Use image-to-video from the reference frame for every shot that includes the character. When something drifts, regenerate the reference rather than the whole sequence.

Lock the palette early

Choose three to five colour values for the film and apply them through both production design and grading. Unrelated shots that share a palette feel like one world; technically perfect shots with five different colour temperatures feel like a showreel. The same logic applies to grain and contrast: a single film grain pass over everything, generated and filmed alike, does more for perceived quality than a fourth generation attempt on one stubborn shot.

A worked example: the 45-second workshop film

Take a furniture maker with a small workshop and a strong story. Six beats, roughly seven seconds each: the hook (hands, material, sound), the context (the workshop, the town), the craft (a tool meeting wood), the human (a face, a decision), the proof (the finished piece in a real room), the close (name, one line, one action).

Write each beat as a sentence with a verb in it. Beats become shots. Shots become prompts. Never the other way around.

Generate three variations per shot, not one. Judge them at thumbnail size first, because a shot that reads as a smudge in a contact sheet will read as a smudge on a phone in a feed. Promote the survivors to full size, then to the timeline.

Assemble in this order: picture to temp music, then sound design, then grade, then captions. Grading before sound produces beautiful films with no rhythm. Set a delivery look and apply it to everything, including stills and graphics.

Sound, captions and the finish that sells

Half of perceived production value is audio. A perfectly generated shot with thin sound reads as amateur; a modest shot with a rich sound bed reads as expensive.

Use three layers, always. Ambience is the room: workshop hum, street rain, café chatter, wind on a hillside. Spot effects are the story: the tool strike, the kettle, the door, the fabric. Music is the emotional argument. Duck the music under the most important spot effects, and build one deliberate silence somewhere in the film. Silence is a cinematic tool and it costs nothing.

If you use synthetic narration, write for the ear, not the page: short sentences, concrete nouns, one idea per line. Then listen on a phone speaker, because that is where most of your audience will hear it.

Captions are typography, not an afterthought

Burned-in captions are the default on social, but they must be styled rather than defaulted. High contrast, one line at a time, inside the platform safe zones, never over the product or a face. Keep a subtitle file alongside the master for accessibility and for repurposing into other languages later.

Choosing tools by job, not by hype

Model comparisons age quickly; decision criteria do not. Score each candidate against your actual shot list.

  • Shot-type fit. Does it handle your dominant shot type, whether that is faces, product macro, landscape or stylised motion?
  • Control surface. Image-to-video, start and end frame keying, camera controls and duration limits tell you how much authorship you keep.
  • Cost per usable second. Track how many generations it takes to get one shot you would actually cut. That number is your real cost.
  • Iteration speed. A slightly weaker model that returns results in half the time usually wins on a fixed deadline.
  • Rights and disclosure. Confirm commercial usage terms in writing and decide how you will disclose synthetic footage to clients and audiences.
  • Pipeline fit. Resolution, frame rate and export formats that drop into your edit without conversion save hours per project.

Where a single workspace helps

Most small teams lose time not in generation but in moving files between five browser tabs. Consolidating generation, stills and reusable structures in one workspace shortens the loop between idea and cut. That is the practical argument for a tool like Orelon's AI video generator: fewer handoffs, faster iteration and a consistent output pipeline from first test to final export. If you are weighing options, a direct comparison such as Orelon vs Runway is more useful than a feature table, because it frames the decision around workflow rather than benchmark scores. Reusable video templates matter more than agencies admit: a repeatable structure for a 15-second social cut means a new client project starts well past the blank page.

Review, delivery and the client conversation

A QA checklist that prevents reshoots

  • Watch on a phone, then a laptop, then a TV, in that order. Note where attention drops.
  • Watch the first three seconds with sound off. If the hook fails silently, it fails.
  • Freeze on cut points and check wardrobe, props and light direction for continuity breaks.
  • Inspect faces and hands at full size, which is where generation artefacts hide.
  • Verify brand assets: logo clear space, product accuracy, legal lines, end-card legibility.
  • Check caption timing and safe zones in every aspect ratio you deliver.
  • Confirm the audio never clips and loudness is consistent across the set.

Plan the version matrix before you edit

Sixteen-by-nine master, nine-by-sixteen vertical, one-by-one for feeds, and a six-second bumper. Each needs its own crop logic, not an automatic centre cut. Deliver a high-bitrate master plus compressed platform versions, and keep a textless version for future localisation. Hand over the project structure as part of the package; clients remember it when they need a variant next quarter.

Scope in three lines

Price creative development (brief, shot list, references), production (generation, iteration, licensing of any stock), and finishing (edit, sound, grade, captions, versioning) separately. State the iteration allowance explicitly, for example two rounds of shot-level revisions and one round of structural changes. Clients accept limits that are stated up front and resist open-ended ones. Track hours against your estimate for the first three projects; your real cost per usable second will surprise you, and it belongs in the next quote.

Mistakes that cost the most

Generating before the shot list exists. Judging footage at full size instead of thumbnail size. Letting each shot carry a different prompt structure. Treating sound as a final step. Leaving the vertical crop until the day before delivery. Promising a look the model cannot hold for more than eight seconds, which is a scoping error rather than a tooling one. And forgetting to write down the disclosure plan, which turns a routine delivery into an awkward conversation.

FAQ

Do we need to tell clients that footage is synthetic? Yes, in writing, and preferably in the contract. Explain the split between generated and filmed material, confirm usage rights, and agree how the work will be described publicly. Clients are far more comfortable when the disclosure plan is proactive rather than a surprise discovered in a comment thread.

How many finished shots can a two-person team complete in a week? Usually eight to twelve, including iteration, editing, sound and one delivery format. Each additional aspect ratio adds roughly a day. Budget by finished shot, never by generation attempt.

Should we generate stills first or go straight to video? Stills first, almost always. They are cheap to iterate, easy for clients to approve, and they convert into starting frames that hold character and composition far better than text alone.

What is the biggest predictor of quality? The strength of your reference stills and the specificity of your camera and lighting language. Teams with strong references and a locked shot list consistently beat teams with better tools and vaguer briefs.

How do we keep a place recognisable without turning it into a postcard? Use detail instead of landmarks. A specific gate, wet pavement, a particular shade of stone, a local accent in the voiceover. One authentic detail per shot beats a skyline montage, and it travels better on social because it feels observed rather than borrowed.

Can this replace a camera crew entirely? No, and pitching it that way damages trust. It replaces the shots that would otherwise break the budget: establishing plates, texture inserts, second-unit material and pitch animatics. The performance, the product truth and the story decision stay yours.

Start with one film, not a system

Pick a single client with a clear story and a modest runtime. Write six beats, build twelve reference stills, generate three variations per shot, and finish it properly with sound, captions and two aspect ratios. That one project will teach you more than a month of tool comparison, and it gives you something far more persuasive than a demo reel: a finished film with a real brand on it.

When you are ready to compress the loop between idea and cut, Orelon keeps generation, stills and templates in one place so the work stays cinematic from the first reference frame to the final export. Bring the shot list. The rest gets faster.